基于建模的视觉定位（SFM-Based Positioning）

具体方法来自我参与的这篇journal：

Vision-Based Positioning for Internet-of-Vehicles， IEEE Transactions on Intelligent Transportation Systems, 2016.

过程：基于图像的3D建模 --> 点云压缩 --> 3D-2D的匹配

Method：http://www.clarenceliang.com/positioning

Dataset：http://www.clarenceliang.com/dataset

 

注：以下內容是個人筆記，source code目前還沒有release出來。但其中VisualSFM和2D-3D Matching的部份是公開的，可以在他們的主頁找到。

Introduction

images/:   put your training images in this folder

testImages/:   put your testing images in this folder

bundle/:   put the .out file generated by visualSFM in this folder

file_gen/:   files generated in the compression step

result/:   results generated in the localization step

work_flow_2.m:   do the model compression

BatchLocalizer.sh:   script to do the localization of the test images

simple_test.m:   generate the test result without ground truth 

bash_test.m:    generate the test result with ground truth

Steps to use the code

• Training Phase

1. Install visualSFM

http://ccwu.me/vsfm/

2. Open VisualSFM

3. Load Images

File->Open + Multi Images->select the training images at ‘Positioning/images/’

4. Feature Matching

Click on ‘Compute Missing Matches’ 

5. 3D Reconstruction

Click on ‘Compute 3D Reconstruction’

6.  Re-order the cameras

Hit ENTER-> ‘sort’->ENTER

7. Save Results  

Sfm->Extra Functions->Save Current Model->‘bundle.out’ at ‘Positioning/bundle/’

  Save Current Cameras-> ‘list.txt’ at ‘Positioning/’

8. Close VisualSFM

9. Model Compression

Command line -> ./bin/siftb2a list.txt

Execute ‘work_flow_2.m’ in Matlab，record the value of variablepwk.

• Testing Phase

10. Localization

Command line -> ./BatchLocalizer.sh bundle/bundle.out list.txt file_gen/cluster_k_185.txt 100(10^pwk)testImages/(path of testing images)result/(path of results) 100(testing times) 0.4 100

11.  Show Testing Result

Execute ‘simple_test.m’ in Matlab, get ‘trainM.mat’(positions of training images), ‘testM.mat’(positions of testing images), ‘point_position.mat’(positions of model points), ’point_color.mat’(colors of model points).

12. Compute The Error With Ground Truth(optional)

Write the positions of the ground truth images into a new matrix ‘trainR.mat’. Then only keep the corresponding rows in ‘trainM’ and delete the others. Write the positions of the testing images into a new matrix ‘testR.mat’. 

Execute ‘bash_test.m’ in Matlab, getrefand dev. Variable refstores the position of each test image in the ground truth coordinate.Dev is the error between each test image and the ground truth.